Trustworthy Bayesian deep learning framework for uncertainty quantification and confidence calibration: Application in machinery fault diagnosis

Reliable and accurate machinery fault diagnosis is crucial for ensuring operational safety and reducing downtime in industrial settings. Traditional intelligent diagnosis methods only focus on improving the accuracy of in-distribution samples, but neglect the trustworthiness evaluation of diagnosis results. To address these issues, this paper developed a novel trustworthy machinery fault diagnosis (TMFD) method, which integrates Bayesian deep learning techniques with model calibration strategies. Specifically, TMFD regards a Bayesian convolutional neural network framework as the backbone. Then, we introduce α-divergence to facilitate the decomposition and quantification of epistemic uncertainty and aleatoric uncertainty, ultimately achieving out-of-distribution sample detection through epistemic uncertainty. Then, the ante-calibration loss constraint and the compositional post-calibration operation are jointly applied to promote data-efficient and high expressive calibration for in-distribution sample diagnosis confidence. Finally, TMFD is validated using three experimental datasets, demonstrating its effectiveness and robustness in machinery fault diagnosis.